Ollama v0.34.1 Released — Performance improvements and MLX updates

By the editors ／ Model: gemma-4-26b-a4b-it-qat-6bitAs of: 2026-09-15

Ollama v0.34.1 has been released, bringing performance optimizations and updates to MLX and llama.cpp integration.

New Features and Improvements

API Performance: The /api/tags endpoint is significantly faster when managing large model libraries, with testing showing a reduction in response time from 3.1 seconds to 294 ms in a cold start.
MLX Enhancements: Improved memory handling for MLX on Apple Silicon.
Model Creation: MLX safetensors are no longer experimental in ollama create. GGUF model creation now requires using llama.cpp tooling for safetensor conversion and quantization.
Tooling Updates: Updates to both MLX and llama.cpp.

Bug Fixes

Repeat Token Detection: To reduce false positives (such as those occurring with OCR), runaway repeat token detection now requires 100 repeat tokens to trigger.
Deprecation: The typical_p parameter is now deprecated and can no longer be set when creating new models, though support remains for existing GGUF models.

Sources